Daily Brief Nov 13, 2026 · ← Back to this issue
How we checked every item today except "Also happened"
For readers who want to dig in. For what happened, in brief, go back to this issue.
This issue draws on material from November 12 to early November 13. Last night's collection took in 446 pieces, 109 of them read in full, and this issue uses 23 outside sources with checkable links; the later sections also draw on material from the past week. The full edition of this issue is the archive of record.
1. [Today] NVIDIA's CFO Colette Kress becomes acting CEO, and the November 17 earnings report stays on schedule
Not proven yet： the third quarter's end date is SecondSource's calculation; the filing doesn't state it
How we checked
The resignation and the acting appointment come from the current report NVIDIA filed with the SEC, a legally required company filing. The permanent CEO search and the November 17 results date come from NVIDIA's press release. The fiscal-year rule, payment terms, accounts receivable, investment commitments, and the guarantee caps and timing come from NVIDIA's second-quarter report, which was reviewed by the auditor but not audited. The current report doesn't address the removal process the media reported earlier, and the board's meeting minutes aren't among the public materials. Neither NVIDIA nor Kress has said publicly whether Kress will change payment terms or investments. The October 25 date is SecondSource's count. NVIDIA's second-quarter report says the fiscal year ends on the last Sunday of January. This fiscal year has 53 weeks with a 14-week fourth quarter, so the first three quarters run 13 weeks each. The second quarter ended on July 26, and 13 weeks later is October 25 (our calculation; NVIDIA second-quarter report, 2026-08-26).
2. [Today] Google launches the low-cost model Gemini 4 Hydrogen, which scores 57% on the Terminal-Bench 4.1 terminal-task benchmark in Google's own test, against 18.2% for the previous Gemini 3.8 Flash
Not proven yet： Google's same-class comparison names no other company's model
How we checked
Everything above comes from Google DeepMind's announcement. Google says Hydrogen leads its class but doesn't name which companies' models, or which versions, it compared against; the only scored comparisons are Google's own Gemini 3.8 Flash and the larger Gemini 4 Argon. According to Google, on the Artificial Analysis Intelligence Index, a third-party index of overall model capability, Hydrogen scores 51 at default settings and Argon 50; these figures haven't been checked against the index itself. Google says Hydrogen completes tasks 2.5 times as fast as Gemini 3.8 Flash and that partners' early tests found it 2 to 2.8 times as fast; neither claim states the task set or sample.
3. [Today] Perplexity releases the classification model Decider, which it says costs about 1/12 as much as GLM-5.3-Flash run through an API, not counting local hardware
Not proven yet： the reference answers for accuracy came from a frontier-model judge
How we checked
This comes from Perplexity's announcement and the Hugging Face model page (Hugging Face, 2026-11). Every comparison with figures is an evaluation Perplexity built; the live-traffic set comes from about 2.32 million questions over the 18 days before launch. For three outside leaderboards, from Hugging Face, Vals AI and Artificial Analysis, the announcement says only that performance is strong and gives no scores. The model page says v1.3 has 26 billion parameters and is built on Qwen3.8-27B, and doesn't explain the difference between the two numbers. The model page also says the top answer changes 1.5% to 2.2% of the time, depending on which other questions are sent in the same request.
4. [Today] AI cloud provider Nebius forecasts 2027 capital spending of US$40 billion to US$45 billion, while 2026 revenue lands at the low end of guidance
Not proven yet： the figures are self-reported by Nebius, and the shareholder letter isn't audited
How we checked
Nebius wrote all of the above itself, in two current reports filed with the SEC as a foreign issuer that contain the earnings release, the shareholder letter and the operating review. The 1.5 to 1.7 times is SecondSource's calculation, dividing US$40 billion and US$45 billion by US$27 billion. Nebius also writes that it has more than 4GW of contracted power; that is capacity under signed power or site contracts, not capacity already online. The shareholder letter doesn't say how much is online now.
[Today] Applied Materials says revenue at its semiconductor equipment segment will grow more than 40% over a one-year period the company defined, running from the second quarter of fiscal 2026 through the first quarter of fiscal 2027.
How we checked
This is one company's segment revenue, not the size of the global wafer equipment market, and the press release doesn't say which period it is compared with.
[Today] TSMC filed with the US SEC that data on some current and former employees appeared without authorization on an outside website and matches data in a cloud service provider's systems.
How we checked
TSMC writes that it has found no evidence so far that TSMC's own systems were breached, and the filing doesn't give the number of people affected or the provider's name. TSMC writes that it has contacted and is working with the cloud service provider, the relevant investigative authorities and a third-party cybersecurity firm, and is notifying the people affected.
[This week] Epoch AI, an independent research group that studies AI compute, estimates from satellite imagery that Huawei's AI campus in Horinger, Inner Mongolia, holds about 81,000 H100 equivalents of compute, about 7% of the compute at Colossus 2, an AI data-center campus in Tennessee.
How we checked
This is Epoch's model estimate, with a 90% interval of 36,000 to 141,000, and it assumes the campus holds no export-controlled NVIDIA chips. By Epoch's estimate, this campus gets about 40% of the compute per unit of power that leading US campuses get. Combining that with its own estimate of Huawei chip sales, it puts about 105,000 Ascend 910C chips in operation (Epoch AI, 2026-11-06).
[Today] Chip research firm SemiAnalysis argues that NVIDIA is a company built around Jensen Huang with a thin succession bench, and that the two internal technical leaders being talked about would struggle to take over Huang's command of the supply chain.
How we checked
SemiAnalysis attaches no data to this judgment, and it sells data to chipmakers and investment firms.
[Today] Eli Lilly CEO Dave Ricks, responding to Anthropic CEO Dario Amodei's claim that major diseases will be a thing of the past within ten years, says that would be great but is unlikely.
How we checked
Amodei's remark comes from a video connected with Anthropic's IPO, reposted by a third-party account. Marcus doesn't say whether the IPO registration statement contains the same words.
[This week] A preprint finds that a single fake news article in an AI forecasting system's search results can reverse 74% to 82% of the forecasts that originally opposed the article's position, while average accuracy barely changes.
How we checked
According to the paper's abstract, using two retrieval-side defenses together changes 29% of normal judgments and discards 39% of genuine news, and these are the authors' self-reported results. No one else has repeated the results (arXiv 2611.15102, 2026-11-11).
[This week] The preprint Lazarus plants a backdoor in open-weight agent models by rewriting just 7 neurons, without retraining, a result the authors self-report.
How we checked
[Today] Google tested its tool Agent Anomaly Detection, which flags agents overstepping after the fact, on agent-safety benchmarks written by 5 outside teams. The tool flagged 87% of harmful instructions that were executed, but caught only 4 of 18 poisoned instructions written into long-term memory.
How we checked
[Today] Databricks' data agent Genie is now generally available in Slack, querying data with the asker's own permissions and requiring user approval before changing outside documents, with no usage or pricing figures published.
How we checked
Still missing today. The gap that most affects judgment is the end date of NVIDIA's third quarter: the filing doesn't give it, and this issue uses SecondSource's calculation. The original documents behind Lam Research's and KLA's estimates of 2027 global wafer-equipment spending haven't been read. The material cited in this issue includes no independent comparison of Gemini 4 Hydrogen with other companies' models. Anthropic's IPO price range rests only on Zvi Mowshowitz's citation and hasn't been checked against the registration statement. The original posts and reports behind the four Also happened items haven't been checked word for word. Of the 216 pieces not yet read, 104 are academic papers and 69 are social-platform posts.
Sources & accounting
The past 24 hours. 446 new pieces came in from November 12 to November 13: 183 academic papers, 151 social-platform posts, 69 company and personal blog posts (including Taiwan Stock Exchange material-information filings), 20 linked articles, 10 subscription newsletters, 8 SEC filings and 5 from a second academic-paper feed (counted separately from the academic papers). Of those 446, 109 were read in full, 111 were screened out as off-topic (5 of them before fetching), 9 were only indexed without a full read, 1 is waiting to be refetched, and 216 haven't been read yet. Academic papers, 183: 23 read, 56 screened out, 104 not yet read. Social-platform posts, 151: 40 read, 41 screened out, 1 only indexed, 69 not yet read. Blogs, 69: 17 read, 5 screened out, 8 only indexed, 1 waiting to be refetched, 38 not yet read. Linked articles, 20: 17 read, 3 screened out (2 of them before fetching). Newsletters, 10: 8 read, 2 screened out. SEC filings, 8: 4 read, 1 screened out, 3 not yet read. Second paper feed, 5: 3 screened out (all 3 before fetching), 2 not yet read. Last night no fetch failure was reported for arXiv, blog feeds, X or SEC filings; the podcast source check failed, and no podcast material is used in this issue's body. This window had no new macro data.
One-time catch-up. No new one-time catch-up material came in this window. The catch-up material already collected is dated between July 1 and November 9 this year, counted separately from the past 24 hours above, and none of it was used in this issue.
Source concentration. Main-line item 1, Named commentary item 1 and Also happened item 3 all concern NVIDIA's succession. Main-line items 2 and 3 and both Product moves items are the companies' own announcements, with mostly self-tested figures. All four Also happened items are relays, and the original posts and reports haven't been checked for this issue.
Tracked sources. The long-term roster currently has 529 named sources: 302 on social platforms, 90 shows, 51 news outlets, 48 blogs, 48 paper authors and 46 newsletters, with the rest spread across earnings reports, keynotes and other channels. The same source may be tracked on several channels, so the channel figures add up to more than 529. These are counts of tracked sources, a different list from last night's piece counts above, so the two can't be added or compared. Representative names include Dan Nystedt, Ming-Chi Kuo, Paul Triolo and Joe Weisenthal on social platforms; Zvi Mowshowitz, Gary Marcus and Dylan Patel among newsletters; and Lisa Su and Satya Nadella on the earnings roster. This issue uses 23 outside sources in the body. Only outside links actually cited in the body are counted. Sources from Taiwan total 240, with 43 added in October; Taiwanese media, Taiwanese companies' official channels and local reporters in Taiwan count, while foreign outlets relaying Taiwanese media do not.